Proceedings of the 6th International Conference on Computing and Informatics, ICOCI 2017
25-27April, 2017 Kuala Lumpur. Universiti Utara Malaysia (http://www.uum.edu.my )
Paper No.
000
MODELLING CONTENTS STATUS FOR IPTV DELIVERY
NETWORKS
Suliman Mohamed Fati1,Putra Sumari2, Siti Sophiayati Yuhaniz3, Nilam
Nur Binti Amir Sjarif4
1Faculty
of Information Technology-Math & Science, Inti Internatioanl University, Persiaran Perdana BBN,
Putra Nilai, 71800 Nilai, Negeri Sembilan
2 Multimedia Research Group, School of computer sciences, Univeristi Sains Malaysia, Malaysia
3,4Advanced Informatics School, Universiti Teknologi Malaysia,54100 Kuala Lumpur, Malaysia.
1
[email protected], [email protected], [email protected],[email protected]
ABSTRACT. Since IPTV has been invented, IPTV is considered a dominant technology to distribute high quality videos and live channels anytime
anywhere over challenging environment to end users who are having different preferences and demands. Presently, IPTV service providers manage
IPTV delivery networks, in terms of contents, channels, resources, based on
contents popularity distribution and/or users’ preferences only. Although
content popularity and users’ preferences play an important role to cope
with the increasing demand of IPTV contents/channels, these two measures
fail in producing efficient IPTV delivery networks. For that, IPTV delivery
network designing should integrate the IPTV content characteristics like
size, interactivity, the rapid changing lifetime. Therefore, the idea of this
paper is to build a mathematical model that integrates all these factors in one
concept called IPTV content status. Modeling the contents status according
to its characteristics is an important point to design Content-Aware IPTV
delivery networks. The experimental results showed the superiority of modeling IPTV content status in balancing the load and reducing the resources
waste.
Keywords: IPTV, Delivery Networks, Content Popularity, Contents’ workload, IPTV content status.
INTRODUCTION
With the massive development in networking and multimedia fields, IPTV has become recently popular as a promising trend in home and business entertainment due to the increasing
number of IP-networks users. Therefore, IPTV, as a promising technology, grows rapidly in
terms of subscribers and revenue to become, in the near future, the standard means to deliver
home and business entertainment contents (Yarali & Cherry, 2005). Thus, Telecommunication companies have entered a hectic competition to increase their customer base and profit
by providing IPTV services (Yarali & Cherry, 2005) (Lee, et al, 2009) (Li & Wu, 2010).
Nowadays, IPTV is globally considered a dominant technology to distribute high quality videos and live channels anytime anywhere over challenging environment. This challenging environment tends to deliver a hybrid IPTV services over IP-based network infrastructure into
end-users with different preferences and demands (Aldana Diaz & Huh, 2011) (Montpetit,
Klym & Blain, 2010) (Montpetit, Klym, & Mirlacher, 2011).
Proceedings of the 6th International Conference on Computing and Informatics, ICOCI 2017
25-27April, 2017 Kuala Lumpur. Universiti Utara Malaysia (http://www.uum.edu.my )
Paper No.
000
Currently, IPTV service providers consider the IPTV contents popularity distribution
and/or users’ preferences to allocate IPTV contents (Dukes & Jones, 2004) (Sujatha, et al,
2008) (Lie, Lui, & Golubchik, 2000) (Little & Venkatesh, 1993) (Sobe, Elmenreich, &
Böszörmenyi, 2010), balance the network workload (Huang & Fang, 2004) (Zhou & Xu,
2002), and/or pre-fetch the desired channels into the Set-top-box (Bikfalvi, 2011). Moreover,
IPTV service providers exploit the users’ preferences and content popularity to recommend
some contents and also to control the advertisement that should be displayed to the users.
Although content popularity and users’ preferences play an important role to cope with the
increasing demand of IPTV contents/channels, these two measures fail in producing efficient
IPTV delivery networks (Fati, Sumari, & Budiartu, 2015) (Fati, Budiartu, & Sumari, 2014)
(Fati & Sumari, 2016). For example, allocating the IPTV contents according to the content
popularity only may lead to satisfy the users’ demand or preferences. According to (Joe, Yi &
Sohn, 2011), the popularity-based replication schemes assume a threshold to decide which
contents should be replicated. In other words, the contents having popularity values larger
than this threshold can be replicated. Also, some of these algorithms are proactive wherein the
contents can be replicated once the server allocating these contents becomes overloaded. Actually, replicating the huge size contents, like IPTV contents, should consider the trade-off
between the QoS requirement and exploited storage space. In other words, the replication
algorithms should maintain the QoS requirements without wasting the storage resources. Both
the popularity based and proactive algorithms may replicate redundant contents, which leads
in turn, to increase the exploited storage space. On the other hand, the current studies focus on
the channel popularity and/or user preferences only. In the real world, the popularity of channel or the user behaviour suffer from the rapid fluctuation. Such fluctuation in channel popularity and/or users’ preferences may lead to prefetching irrelevant channels, which in turn,
leads to bandwidth saturation and delay the required channel, as well.
To overcome this challenge, the design process of IPTV delivery networks for both Ondemand and live services should consider the other characteristics of IPTV contents/channels.
Beside the popularity, these characteristics include the huge size (e.g. HD requires huge
bandwidth more than the normal channels), interactivity, the rapidly changing effective (Joe,
Yi & Sohn, 2011). In addition, there are other factors should be considered including multicast factor (e.g. some channels are requested by many people in the area while other channels requested rarely or by individuals. The population size is another factor should be considered (e.g. some service areas are crowded and then the request rate is higher than other
areas.
According to (García, et al, 2009) (Fati & Sumari, 2016), modeling the workload for multimedia streaming services is essential factor in delivery networks to improve the performance, enhance the QoS, and increase the reliability. Therefore, the idea of this paper is to
build a mathematical model that integrates all these factors in one concept called IPTV content status. Modeling the contents status according to its characteristics is an important point
to design Content-Aware IPTV delivery networks. Content status modeling refers as estimating the portion of concurrent requests that target the content according to its characteristics.
Such modeling helps a lot in handling both the load balancing and resource allocation based
on the anticipated load of contents. However, these contents are huge size with non-uniform
patterns of both contents lifetime and users request access. For this reason, modeling the contents status is a challenge. Many works in the literature review focus on the popularity of contents during their approaches to solve load imbalance problem. However, content popularity
is not everything. The other characteristics like the size and the interactivity play a significant
role in modeling the content status. Hence, the aim of this paper is how to model the content
status in IPTV environment. Modeling the content status helps the delivery networks provid-
Proceedings of the 6th International Conference on Computing and Informatics, ICOCI 2017
25-27April, 2017 Kuala Lumpur. Universiti Utara Malaysia (http://www.uum.edu.my )
Paper No.
000
ers to build content-aware delivery networks to overcome the load imbalance and resources
waste problems. To the best of our knowledge, there is no work formulate the content status
for IPTV contents.
2.
BACKGROUND AND RELATED WORK
The expected load of content or server has a significant role in balancing the load among
the servers. Estimating the load of content or server takes different forms according to the
nature of problems. In some works, the load can be estimated as the CPU consumption while,
in other works, the load can be estimated as the active connections to the servers (Cho, et al,
2008) (Meng, Liu, & Yin, 2013) (Kanrar, 2012). In the present paper, the load is estimated as
the number of active concurrent requests that access the content/servers. Thus, the IPTV content status model that estimates the maximum load for each video in the service area at the
peak busy period is suggested. The video load is estimated based on its popularity, length, and
request arrival rate (e.g. normal requests and interactive requests) (Gaber & Sumari, 2014)
(Fati, Budiartu, & Sumari, 2014). The proposed estimation model deals with both the normal
and interactive viewing sessions. The load and the workload will be used interchangeably
throughout this thesis.
For each video, the workload can be estimated as the number of concurrent requests that
must be served by that video. The estimated workload for a video is a function of the request
arrival rate, the number of subscribers within the service area, the video popularity, and the
video length. All these information, which are required in the workload estimation, should be
stored in the request dispatcher that is located in front of the clustered VHO. To estimate the
workload in IPTV system, we have to imagine the viewing scenario of any subscriber as follows: Once connected, the subscriber has two viewing states; a normal state and an interactive
state. The user starts in the normal state where the requested video is being viewed at the usual speed. After that, the subscriber issues an interactive VCR command to stop, pause, and
jump forward or backward (Cho, et al, 2008). During the viewing time, the user can transit
from one state to another interchangeably until the end of viewing time.
Based on the above viewing scenario, there are two types of requests: normal requests and
interactive requests. Thus, the expected workload of a video i can be defined as the portion of
concurrent requests of both types that can be served by this content within the peak busy period. Therefore, the expected load of content i is estimated as the fraction of simultaneous incoming requests targeting content i.
According to (Kanrar, 2012) (Wong, 2004), the number of concurrent requests at a certain
time point can be estimated as the sum of the active requests lengths within an interval time
divided by this interval time. Kanrar (Kanrar, 2012) has estimated the server’s workload as a
function of number of households, request arrival rate, and video length during the peak busy
period.
Most of studies focus on the channel popularity and/or user preferences only without any
consideration for the channel characteristics like size (e.g. HD requires huge bandwidth more
than the normal channels), multicast factor (e.g. some channels are requested by many people
in the area while other channels requested rarely or by individuals, and the population size
(e.g. some service areas are crowded and then the request rate is higher than other areas.
Moreover, the popularity of channel or the user behavior IPTV systems suffer from the rapid
fluctuation of users’ preferences in the real world. Therefore, such fluctuation in channel popularity and/or users’ preferences may lead to prefetching irrelevant channels, which in turn,
leads to bandwidth saturation and delay the required channel, as well. For that, expecting the
channel status (the expected requests targeting this channel) according to the channel charac-
Proceedings of the 6th International Conference on Computing and Informatics, ICOCI 2017
25-27April, 2017 Kuala Lumpur. Universiti Utara Malaysia (http://www.uum.edu.my )
Paper No.
000
teristics and demand is a beneficial. This prediction can control the fluctuation in the channel
popularity and/or the users’ preferences. The channel status can be modeled as a function of
channel bandwidth, channel popularity, multicast factor, and region population. According to
the history analysis, we can easily, predict the demand of this channel in a particular area,
then, decide the prefetching.
3.
IPTV CONTENT STATUS MODEL
As mentioned above, the workload of contents can be estimated by considering the content’s characteristic. For example, the content workload means the number of active requests
targeting that content at the same time during the peak busy period. In the case of estimating
the workload for individual content, there is a population of subscribers H issuing different
requests targeting diverse contents at the peak busy period. Each subscriber can issue a number of normal requests λ and interactive request λvcr with holding times t i and t vcr for normal and interactive requests, respectively. Thus, the workload for a video i can be estimated
as the portion of concurrent normal and interactive requests originated within the service area
by the subscribers to watch that content during the peak busy period. This can be interpreted
mathematically as in equation 1.
Li =
pi ∗ H
Tpeak
((λ ∗ t i ) + (λvcr ∗ t vcr ) )
(1)
Where the terms Li , pi and t i denote the expected work load, the popularity and the length
of the particualr movie i in the service area, respectively. The popularity takes a value between 0 and 1. The term H denotes the number of households in the service area, and Tpeak
denotes the peak busy period of IPTV system in minutes. In addition, the terms λ and λvcr
denote the request arrival rate and VCR commands request arrival rate, respectively. Both
arrival rates follow Poisson distribution. In equation (1), the term H((λ ∗ t i ) + (λvcr ∗ t vcr ) )
represents the total sum of all requests issued by the household s within the peak busy period.
By multiplying this term by the content popularity value pi , the number of requests targeting
this content is obtained. Finally, the concurrent requests per a unit time are estimated by dividing on the term Tpeak .
After contents workload estimation, the expected workload of server j is estimated by
summing up the load of all contents that are stored in this server as shown in equation 2.
Lj = ∑i∈j Li =
H
Tpeak
∑i∈j(p ((λ ∗ t i ) + (λvcr ∗ t vcr )))
i
(2)
The term Lj denotes the expected workload for a particular server j. Assuming that the
available servers in the service area are sufficient to handle the total workload, thus, these
servers must be allocated accurately to cope with the growing load problem. After estimating
the expected load of contents, the replication degree (i.e. the number of replicas that are required to handle the expected load) is calculated. The replication degree must be controlled by
the video popularity. For instance, video i can be allocated on all servers if its popularity is
extremely high. This is to ensure that the expected load can be distributed on as a large number of servers as possible to minimize the request rejection rate. On the other hand, there is no
need to replicate the remarkably low popular videos so that one copy is enough to catch its
low expected load. Based on the above explanation, the replication degree (i.e. number of
copies) for a video i can be formulated as a function of the number of servers and the normalized popularity as shown in equation (3). According to (Nafaa, Murphy & Murphy, 2008),
normalizing the popularity distribution improve the overall performance by building a strong
relationship between the content popularity and the replication degree.
Proceedings of the 6th International Conference on Computing and Informatics, ICOCI 2017
25-27April, 2017 Kuala Lumpur. Universiti Utara Malaysia (http://www.uum.edu.my )
ri = ⌈Sa ∗ pni ⌉
Paper No.
000
(3)
Where the term ri denotes the number of replicas for content i , Sa represents the number
of servers in service area a , and finally the term pni representat the normalized value of pi
that should be within [0,1]. The normlized popularity value has been obtained using Min-Max
Normalization Law. Min-max normalization performs a linear transformation on the original
data (Gaber & Sumari, 2014). Min-max normalization maps pi to pni value in the
range [new_min, new_max]. In this case, the new_max value equals one where the highest
popularity value in the dataset will be scaled to equal the value (1) and the other values will
be scaled successively. The operator ⌈. ⌉ is a ceiling function operator to take the largest integer nearest to the calculated term.
After computing the expected number of replicas for a video i, the expected load for each
replica can be calculated by dividing the expected load for that video on its number of replicas as follows:
Lri = Li /ri
(4)
Where Lri represents the load of one replica for content i, Li denotes the load of video I,
and ri represents the number of replicas for content i, which obtained from equation. (3).
According to the proposed content status model, the content of high popularity value,
which has a highly expected load, is replicated more to handle the increasing demands. On the
other hand, the contents of low popularity value can be replicated as less as possible for serving the low expected load. Furthermore, allocating the contents and distributing the incoming
requests according to this proposed IPTV content status model can be useful in maintaining
the load as balanced as possible within the service area.
4.
EXPERIMENTAL RESULTS
To investigate the performance of our proposed IPTV content status model, we have tested
the model on an empirical data set that is sampled according to the content popularity distribution. The popularity distributions used are Zipf’s like distribution (Zipf). There are a set of
assumptions that taken into account during the experimental study of the proposed model as
following: the population of subscribers H=1000, each subscriber can issue a number of normal requests λ=2 request/user/minutes and interactive request λ_vcr=2 request/user/minutes
with holding times t_vcr=10 seconds for interactive requests. These values are typical empirical data and widely used by different studies (Cho, et al, 2008). The content size distribution
has been generated randomly between 50-1000 MB. Figure 1 depicts a sample of the content
size distribution in MB.
Figure 1: Random Contents size Distribution
Proceedings of the 6th International Conference on Computing and Informatics, ICOCI 2017
25-27April, 2017 Kuala Lumpur. Universiti Utara Malaysia (http://www.uum.edu.my )
Paper No.
000
The popularity of contents is computed according to Zipf’s Law (Li & Wu, 2010) as in
Equation 5 where i, N, and θ represent the content rank, the total number of contents, and the
skewness degree respectively. The content rank refers to the position in of content in the sorted contents list. The sorted contents list is sorted according to their popularity values. In this
list, the first video is the highest popular video, and so on.
P(i, θ, N) =
1⁄
iθ
1
∑N
n=1( ⁄nθ )
(5)
Then, the workload is computed using our proposed IPTV content status model. Note that
the expected load of server j can be expressed by summing the load of all contents stored in
this server. Figure 2 depicts the expected load for the contents, which is obtained from the
IPTV content status model.
Figure 2: The distribution of expected content load
Figure 3 plots the relationship between the expected workload of contents and the popularity distribution. What is apparent is the fact that the expected load of contents differs from the
popularity distribution. This is due to the other factors that contribute in the workload estimation like video length and interactivity.
Figure 3: The relation between expected load and popularity
Another experiment is conducted on a delivery network with 4 service areas. Each service
area contains four servers. We changed the popularity skewness at each service area such that
each area allocates different replicas for the contents. After that, our request distribution algorithm (Gaber & Sumari, 2014) has been developed over the IPTV content status model. Then,
this algorithm (CARDA) has been compared with other two request distribution algorithms,
Chu (Cho, et al, 2008) and Round Robin (RR) algorithm. The experiment showed that the
popularity has no effect on the request distribution mechanism that supports the argument in
(Huang, et al, 2005). Chengdu (Huang, et al, 2005) argued that the request distribution process is solely affected by the request arrival rates. Figure 4 depicts the distribution of workload among four service areas with four servers for each. Moreover, it is apparent the facts
that our model, which exploit the content status, has a balanced workload distribution among
all servers.
Proceedings of the 6th International Conference on Computing and Informatics, ICOCI 2017
25-27April, 2017 Kuala Lumpur. Universiti Utara Malaysia (http://www.uum.edu.my )
Paper No.
000
Figure 4: The relation between expected load and popularity
5.
CONCLUSION
Managing IPTV delivery networks that allocate interactive huge size contents (video
and/or channels) characterized by fluctuated popularity and non-uniform request patterns is
challengeable. Ignoring such characteristics lead to load imbalance and affects the system
performance. Therefore, modelling the contents status according to its characteristics is very
important in designing IPTV delivery networks. Such modelling helps a lot in handling the
load balancing by distributing the incoming requests based on the anticipated load of both
contents and servers. However, modelling the contents status is challengeable due to the nature of the IPTV contents. Most of the proposed works focus only on content popularity, but
the influence of other characteristics should be investigated. Thus, this thesis suggested the
IPTV content status model that estimates the maximum workload for each video. The load is
estimated as the number of active concurrent requests that access the content. The workload
for the content is estimated in the service area at the peak busy period based on its popularity,
length, and request arrival rate (e.g. normal requests and interactive requests). The proposed
model deals with both the normal and interactive viewing sessions.
REFERENCES
Aldana Diaz, M. E., & Huh, E. N. (2011, February). Cost analysis on IPTV hosting service for 3rd
party providers. In Proceedings of the 5th International Conference on Ubiquitous Information
Management and Communication (p. 114). ACM.
Montpetit, M. J., Klym, N., & Blain, E. (2010). The future of mobile TV: when mobile TV meets the
internet and social networking. In Mobile TV: Customizing content and experience (pp. 305326). Springer London.
Montpetit, M. J., Klym, N., & Mirlacher, T. (2011). The future of IPTV. Multimedia Tools and Applications, 53(3), 519-532.
Yarali, A., & Cherry, A. (2005, November). Internet protocol television (IPTV). In TENCON 2005
2005 IEEE Region 10 (pp. 1-6). IEEE.
Lee, S. B., Muntean, G. M., & Smeaton, A. F. (2009). Performance-aware replication of distributed
pre-recorded IPTV content. IEEE Transactions on Broadcasting, 55(2), 516-526.
Li, M., & Wu, C. H. (2010). A cost-effective resource allocation and management scheme for content
networks supporting IPTV services. Computer Communications, 33(1), 83-91.
Dukes, J., & Jones, J. (2004, November). Using dynamic replication to manage service availability in a
multimedia server cluster. In International Workshop on Multimedia Interactive Protocols and
Systems (pp. 194-205). Springer Berlin Heidelberg.
Proceedings of the 6th International Conference on Computing and Informatics, ICOCI 2017
25-27April, 2017 Kuala Lumpur. Universiti Utara Malaysia (http://www.uum.edu.my )
Paper No.
000
Sujatha, D. N., Girish, K., Venugopal, K. R., & Patnaik, L. M. (2008, January). An efficient storage
mechanism to distribute disk load in a VoD server. In International Conference on Distributed
Computing and Networking (pp. 478-483). Springer Berlin Heidelberg.
Lie, P. W., Lui, J. C., & Golubchik, L. (2000). Threshold-based dynamic replication in large-scale
video-on-demand systems. In Continuous Media Databases (pp. 31-58). Springer US.
Little, T. D., & Venkatesh, D. (1993, November). Probabilistic assignment of movies to storage devices in a video-on-demand system. In International Workshop on Network and Operating System
Support for Digital Audio and Video (pp. 204-215). Springer Berlin Heidelberg.
Sobe, A., Elmenreich, W., & Böszörmenyi, L. (2010, October). Towards a self-organizing replication
model for non-sequential media access. In Proceedings of the 2010 ACM workshop on Social,
adaptive and personalized multimedia interaction and access (pp. 3-8). ACM.
Huang, Y. F., & Fang, C. C. (2004). Load balancing for clusters of VOD servers. Information Sciences,
164(1), 113-138.
Zhou, X., & Xu, C. Z. (2002, April). Request Redirection and Data Layout for Network Traffic Balancing in Cluster-Based Video-on-Demand Servers. In ipdps.
Bikfalvi, A., García-Reinoso, J., Vidal, I., Valera, F., & Azcorra, A. (2011). P2P vs. IP multicast:
Comparing approaches to IPTV streaming based on TV channel popularity. Computer Networks, 55(6), 1310-1325.
Joe, I., Yi, J. H., & Sohn, K. S. (2011). A content-based caching algorithm for streaming media cache
servers in cdn. In Multimedia, Computer Graphics and Broadcasting (pp. 28-36). Springer Berlin
Heidelberg.
García, R., Pañeda, X. G., Melendi, D., & Garcia, V. (2009). Probabilistic analysis and interdependence discovery in the user interactions of a video news on demand service. Computer Networks,
53(12), 2038-2049.
Cho, D. H., Lee, K. Y., Choi, S. L., Chung, Y. D., Kim, M. H., & Lee, Y. J. (2008). A request distribution method for clustered VOD servers considering buffer sharing effects. Journal of Systems
Architecture, 54(1), 335-346.
Meng, S., Liu, L., & Yin, J. (2013). A collaborative and scalable platform for on-demand IPTV services. IEEE Transactions on Services Computing, 6(3), 358-372.
Kanrar, S. (2012). Analysis and implementation of the large scale video-on-demand system. arXiv
preprint arXiv:1202.5094.
Wong, E., M. (2004). Method for Estimating the Number of Concurrent Users, report. Retrieved from:
http://wenku.baidu.com/view/ 68f1853a5802-16fc700afd74.html.
Nafaa, A., Murphy, S., & Murphy, L. (2008). Analysis of a large-scale VOD architecture for broadband operators: a P2P-based solution. IEEE Communications Magazine, 46(12).
Jain, Y. K., & Bhandare, S. K. (2011). Min max normalization based data perturbation method for
privacy protection. International Journal of Computer & Communication Technology (IJCCT),
2(8), 45-50.
Gaber, S. M. A., & Sumari, P. (2014). Predictive and content-aware load balancing algorithm for peerservice area based IPTV networks. Multimedia tools and applications, 70(3), 1987-2010.
Huang, C., Zhou, G., Abdelzaher, T. F., Son, S. H., & Stankovic, J. A. (2005, December). Load balancing in bounded-latency content distribution. In Real-Time Systems Symposium, 2005. RTSS
2005. 26th IEEE International (pp. 12-pp). IEEE.
Fati, S., Sumari, P., & Budiartu, R. (2015). Fair and popularity based content allocation scheme for
IPTV delivery networks. Int J Comput Appl, 131(5), 21-26.
Proceedings of the 6th International Conference on Computing and Informatics, ICOCI 2017
25-27April, 2017 Kuala Lumpur. Universiti Utara Malaysia (http://www.uum.edu.my )
Paper No.
000
Fati, S. M., Budiartu, R., & Sumari, P. (2014, January). Provisioning virtual IPTV delivery networks
using hybrid genetic algorithm. In Proceedings of the 8th International Conference on Ubiquitous Information Management and Communication (p. 106). ACM.
Fati, S. M., & Sumari, P. (2016). Content-aware replica placement strategy for IPTV services over
peer-service area architecture. Multimedia Tools and Applications, 1-25.